字符串数据转换处理300MB，在Dask中具有30M记录分布式

问题描述

在Node1（4cpu，8GB）上启动dask Scheduler：

dask Scheduler：dask-scheduler --host 0.0.0.0 --port 8786

在Node2（8cpu，32GB）和Node3（8cpu，32GB）上启动工作线程： 达斯工人：

dask-worker tcp://http://xxx.xxx.xxx.xxx:8786 --nanny-port 3000:3004 --worker-port 3100:3104 --dashboard-address :8789

这是我的原型，其中编辑了some_private_processing和some_processing方法：

import glob
import pandas as pd

from dask.distributed import Client

N_CORES = 16
THREADS_PER_WORKER = 2
dask_cluster = Client(
    '127.0.0.1:8786'
)

def get_clean_str1(str1):
    ret_tuple = None,False,True,None,False
    if not str1:
        return ret_tuple
    if string_validators(str1) is not True:
        return ret_tuple

    data = some_processing(str1)
    match_flag = False
    if str1 == data.get('formated_str1'):
        match_flag = True

    private_data = some_private_processing(str1)
    private_match_flag = False
    if str1 == private_data.get('formated_private_str1'):
        private_match_flag = True
    ret_tuple = str1,match_flag,private_str1,private_match_flag
    return ret_tuple

files = [
    'part-00000-abcd.gz.parquet','part-00001-abcd.gz.parquet','part-00002-abcd.gz.parquet',]
print('Starting...')
for idx,each_file in enumerate(files):
    dask_cluster.restart()
    print(f'Processing file {idx}: {each_file}')
    all_str1s_df = pd.read_parquet(
        each_file,engine='pyarrow'
    )
    print(f'Read file {idx}: {each_file}')
    all_str1s_df = dd.from_pandas(all_str1s_df,npartitions=16000)
    print(f'Starting file processing {idx}: {each_file}')
    str1_res_tuple = all_str1s_df.map_partitions(
        lambda part: part.apply(
            lambda x: get_clean_str1(x['str1']),axis=1
        ),Meta=tuple
    )

    (clean_str1,bad_str1_flag,private_match_flag) = zip(*str1_res_tuple)

    all_str1s_df = all_str1s_df.assign(
        clean_str1=pd.Series(clean_str1)
    )
    all_str1s_df = all_str1s_df.assign(
        match_flag=pd.Series(match_flag)
    )
    all_str1s_df = all_str1s_df.assign(
        bad_str1_flag=pd.Series(bad_str1_flag)
    )
    all_str1s_df = all_str1s_df.assign(
        private_str1=pd.Series(private_str1)
    )
    all_str1s_df = all_str1s_df.assign(
        private_match_flag=pd.Series(private_match_flag)
    )
    all_str1s_df = all_str1s_df[
        all_str1s_df['match_flag'] == False
    ]
    all_str1s_df = all_str1s_df.repartition(npartitions=200)
    all_str1s_df.to_csv(
        f'results-str1s-{idx}-*.csv'
    )
    print(f'Finished file {idx}: {each_file}')

此过程耗时超过8个小时，我看到所有数据都仅在Node2或Node3上的一个节点上被处理，而在Node2和Node3上都没有。

需要帮助了解这些见解并弄清楚我在哪里做错了，才能使此简单的数据转换运行8小时以上，但仍未完成。

解决方法

超时增加，内存增加。之后，它开始工作，没有失败并挂起。

timeouts:
  connect: 180s          # time before connecting fails
  tcp: 180s              # time before calling an unresponsive connection dead